[Test] SPU LLVM: Add more clamping, reorder them to be on the result

[elad335]

Remove hacks, and reorder clamping where they should be hopefully with better performance. Please test issues that arise when using approx xfloat on master. They can be fixed with this pr in theory.

[Xcedf]

This PR reintroduces an old issue in Uncharted #4478
and this is for FS instruction only, I did a lots of testing with Xfloat clamping on my custom build, clamping final result is worse than clamping it's operands.

[elad335]

I implemented smart float source detection, please retest.

[Whatcookie]

Needs testing with games that are picky with floating point behavior (RDR, LBP2, NCAA)

As for performance, I'm not sure it's going to be any better, I've tested this before and I found that clamping inputs was faster since llvm can eliminate the clamping in the case where the inputs are constant or the same value is used across multiple instructions.

[RainbowCookie32]

GTA 4 no longer needs Accurate xfloat to stay on the ground with this PR

[xddxd]

Causes vertex explosions on horses in RDR. Accurate xfloat on master does this as well, IIRC it's the FM instruction.

[Xcedf]

FM can cause this issue if approximation method was removed and only clamping is used.

[Bertue]

I don't know if it also has to do with the Dragon's Crown and Odin Sphere games with 2D side scrolling because of the problem of texture clamping issues with better optimization in the CPU and GPU of the graphics I have no idea if it would eliminate that so that it doesn't get those square errors out.🤔